Why AI Can Write Code But Can’t Keep It Running (Yet)

Video summary

Grepr CEO Jad Naous says AI cut a six-month production project to 10 days, but fully running distributed systems remains unsolved.

Grepr CEO Jad Naous tells a16z General Partner Martin Casado that observability can consume 5–20% of infrastructure spending, while Grepr’s telemetry filtering can reduce data costs by 90%. AI has transformed his engineering workflow: using Claude Code, Opus, and parallel agent sessions, Naous shipped a complex data-lake autopromotion feature in 10 days—a project he estimates once took one engineer six months. But writing code is easier than operating it: AI still cannot reliably choose production memory allocations or run a distributed cluster without experiments and human judgment. Grepr instead detects unusual log patterns and feeds those early signals to AI for investigation; ignored warnings in customer environments later became incidents. Naous argues AI can learn debugging workflows, even when production failures are unpredictable, while Casado stresses that operations and security face sprawling real-world states and adversaries. They conclude that AI will advance fastest in constrained digital domains, with real-world applications taking longer and often requiring the environment itself to change. Customers, Naous adds, currently care more about Grepr’s 90% data-cost savings than its newer AI reliability features.

Chapters

1. 0:00From Stanford to founding Grepr: Jad’s path from OpenFlow and AppDynamics to observability
2. 7:23Why observability got so expensive: microservices drove up telemetry as reliability gains flattened
3. 10:15How AI is changing engineering: adaptable engineers gained a 10× edge as coding agents evolved
4. 14:02How AI is changing engineering: distributed-systems complexity still requires human judgment
5. 18:25Jad’s AI stack and workflow: WSL, Claude Code, Opus 5.5, and three or four agent sessions
6. 22:55A six-month project in ten days: Jad built data-lake auto-promotion alone
7. 24:55Can AI reason about operations? Benchmarks guide query tuning, but production experiments remain essential
8. 30:10How Grepr catches incidents early: new log patterns and ignored warnings that became incidents
9. 35:01How Grepr catches incidents early: petabyte-scale log scanning and AI-led query investigation
10. 38:35How Grepr catches incidents early: testing causal chains instead of overprovisioning
11. 42:16Why security is the harder problem: adversaries seek rare states, while operations involves people
12. 45:47Why security is the harder problem: distributed human knowledge and the gap between code and reality
13. 49:05Why security is the harder problem: VC relationships, AI analysis, and the digital-world gap
14. 53:33What comes after code: real-world AI, self-driving’s $170 billion, and Grepr’s Datadog savings

What the full analysis includes

14 chapters · 72 key points · 14 section summaries · mind map

The remaining 69 key points, section summaries and mind map are in the full analysis. Analyze this video to see everything →

This video is 59 min. The Free plan analyzes videos up to 30 min; plans from Plus cover videos up to 2 hours.